Learning Local Distribution for Extremely Efficient Single-Image Super-Resolution

Abstract

Achieving balance between efficiency and performance is a key problem for convolution neural network (CNN)-based single-image super-resolution (SISR) algorithms. Existing methods tend to directly output high-resolution (HR) pixels or residuals to reconstruct the HR image and focus a lot of attention on designing powerful CNN backbones. However, this reconstruction way requires the CNN backbone to have good ability to fit the mapping function from LR pixels to HR pixels, which certainly held these methods back from achieving extreme efficiency and from working in embedded environments. In this work, we propose a novel distribution learning architecture to estimate the local distribution and reconstruct HR pixels by sampling the local distribution with the corresponding 2D coordinates. We also improve the backbone structure to better support the proposed distribution learning architecture. The experimental results demonstrate that the proposed method achieves state-of-the-art performance for extremely efficient SISR and exhibits a good balance between efficiency and performance.

An image zooming method based on the coupling threshold in the wavelet packet transform domain

“…Tian et al [19] proposed an enhanced superresolution group convolutional neural network (ESRGCNN), which combined the features of deep and wide channels and extracted low-frequency information of the image using the correlation of different channels. Wu et al [21] estimated the local distribution network structure (LDRN) by a new distribution learning architecture, and on this basis, sampled the local distribution to reconstruct HR images, and Xue et al [15] proposed a super-resolution method for wavelet-based residual attention network (WRAN). We use the ESRGCNN method, the LDRN method, the WRAN method and the algorithm presented in this paper to perform zooming experiments and to compare the experimental results subjectively and objectively from the zoomed images and the PSNR and SSIM metric values.…”

“…These methods train the image pairs of HR images and LR images to obtain the missing high-frequency information in the LR images [13][14][15][16]. Deep learning as a popular field has received the attention of researchers in various industries, and the image processing technology based on convolutional neural networks has achieved better results in image segmentation [17,18], image zooming [19][20][21] and other aspects. These deep learning-based methods perform well and effectively improve the quality of processed images.…”